The AI Bubble Is Real. So Is the Technology.

I’ll say the unpopular thing first: yes, this is a bubble. The valuations don’t make sense, the economics underneath a lot of these companies don’t close, and somewhere in the next couple of years there is going to be a great deal of upheaval. I’m confident about that in the way you can only be confident about things you’ve watched happen before.

I’ll also say the other unpopular thing, the one that annoys the people who nodded along to the first: the technology is real, and it isn’t going anywhere. A bust doesn’t un-invent a capability. When the dot-com bubble burst, the internet did not politely fold itself back up and go home. It kept eating the world. It just did it without a hundred pets-dot-com billboards along the way.

Both of these are true at the same time, and holding them together is the whole trick.

“A bust doesn’t un-invent a capability.”

The froth is not the thing

Most of what people point to when they say “AI” right now is froth. It’s the marketing, the raised rounds, the breathless predictions that we’re eighteen months from a machine god. Strip that away and what’s left is a genuinely useful new tool: a statistical engine that can read, write, summarize, translate, and pattern-match at a scale and speed no team of humans can touch. That part is not hype. I use it every day. My clients get real, measurable value from it.

The problem is that the froth and the thing look identical from the outside, and when the froth collapses (as froth does) it tends to take the reputation of the underlying thing down with it. We’ve seen this movie. A perfectly real piece of technology gets co-opted by a wave of hype, the hype attracts the snake-oil salesmen, the snake-oil salesmen make promises the technology can’t keep, and when the promises come due, everyone decides the whole category was a scam. It wasn’t. It was just oversold by people who didn’t understand it and didn’t care to.

If you lived through the last two or three of these cycles, you can feel the rhythm of it. The same energy that a few years ago was chasing digital collectibles and “big data will change everything” is now pointed at AI. Some of that energy is building durable things. Most of it is going to evaporate. Your job, if you’re an operator, is to make sure you’re standing on the durable part when the evaporating part evaporates.

Trailblazers die; second-movers win

Here’s a pattern I keep coming back to, because it’s one of the most reliable in business history and almost nobody plans around it: the pioneer usually dies.

The trailblazer is the one who fights all the early battles. They spend the capital educating the market, they take the lawsuits, they hit the regulatory walls nobody knew were there, they burn themselves out proving the thing is possible. And then, exhausted and over-leveraged, they stumble, and the second-mover walks in and buys the assets in a fire sale. The follower gets the technology, the lessons, and the cleared path without paying the pioneer’s tax.

I think we’re going to see exactly this in AI. Some of the names that feel unstoppable today are spending money at a rate that only makes sense if the future arrives precisely on schedule and precisely in their favor. It rarely does. When I think about where I’d want exposure, it’s less “who’s in front today” and more “who is quietly positioned to buy the distressed assets when the front-runners crack.” That’s not investment advice (I’m a consultant, not your financial planner) but it’s the lens I can’t stop looking through.

The economics have to close eventually

The other reason I’m confident about a correction is that a lot of the current business models don’t survive contact with a spreadsheet. Charging by the token works beautifully as long as everyone believes the frontier model is the only game in town and demand only goes up. But the frontier is not the only game in town, and it’s getting less so every month.

A well-built routing layer (something that looks at each request and decides how much horsepower it actually needs) can send the overwhelming majority of queries to cheap, small, often local models and only escalate the genuinely hard ones. In practice that can mean only a sliver of your traffic ever touches an expensive frontier model at all. When that becomes normal, and it’s becoming normal fast, the premium-metered-everything business model comes under enormous pressure. Prices fall. The companies whose entire story depended on those prices staying high have a bad year.

That’s healthy, by the way. It’s the market doing its job, moving a genuinely useful commodity toward its real cost.

What actually survives a bust

If you want to know what to hold onto through a correction, it helps to look at what survived the last one. When the internet bubble popped, the froth vanished (the vanity domains, the companies with no revenue and a sock-puppet mascot) but the real things didn’t. The infrastructure stayed. The genuine change in how people behaved stayed. The businesses that had quietly built durable value on top of the real technology didn’t just survive; they inherited the whole landscape once the noise cleared out.

The same filter applies now, and it’s a simple test you can run on any AI initiative, including your own: does this still make sense if the hype completely disappears and the underlying models get ten times cheaper? If the answer is yes (if the thing delivers real value on its own merits regardless of the narrative) it’s durable, and a bust barely touches it. If the only reason it exists is that “AI” is currently a magic word that raises money and moves stock, it’s froth, and it will evaporate exactly when the word stops being magic.

What survives, concretely, is the useful automation woven so deeply into a company’s workflow that pulling it out would hurt. The proprietary data nobody else has. The encoded process that makes your people faster in a way a competitor can’t copy from a press release. None of that depends on the market mood. It depends on whether you built something real. The operators who treated this era as a chance to build durable capability will walk out of the correction stronger, holding assets that still work. The ones who treated it as a story to tell investors will walk out holding a story nobody’s buying anymore. Build the part that survives the pop, and the pop becomes someone else’s problem.

What to do while you wait for the pop

None of this is a reason to sit it out. It’s a reason to be deliberate.

Don’t bet your company on any single vendor’s survival, especially not the flashiest one. Build so that your value lives in your own workflows, your own data, your own encoded way of working, not in a dependency you can’t replace. Assume the model layer is going to commoditize and get cheaper, and design as if that’s a feature, because for you it is.

And honestly? Watch your language. There’s going to come a moment when the word “AI” carries the stink of the bust, the way “dot-com” did for a while. The savvy operators will quietly stop saying it. They’ll talk about “intelligent systems,” or “digital coworkers,” or just describe what the thing does without the loaded acronym. That’s the magic of marketing: the capability stays exactly the same, and only the label gets a fresh coat of paint. I say this half as a joke, but only half.

The people who come out of the next few years ahead won’t be the ones who were loudest during the mania or the ones who were smuggest during the crash. They’ll be the ones who understood, the entire time, that a bubble and a revolution can share a room. They kept building on the part that was real and let the rest float away.

That’s the plan, anyway. Build on the real thing. Ignore the billboards.